| 1 | /* |
| 2 | * Copyright 2019 Advanced Micro Devices, Inc. |
| 3 | * |
| 4 | * Permission is hereby granted, free of charge, to any person obtaining a |
| 5 | * copy of this software and associated documentation files (the "Software"), |
| 6 | * to deal in the Software without restriction, including without limitation |
| 7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
| 8 | * and/or sell copies of the Software, and to permit persons to whom the |
| 9 | * Software is furnished to do so, subject to the following conditions: |
| 10 | * |
| 11 | * The above copyright notice and this permission notice shall be included in |
| 12 | * all copies or substantial portions of the Software. |
| 13 | * |
| 14 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 15 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 16 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 17 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR |
| 18 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, |
| 19 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR |
| 20 | * OTHER DEALINGS IN THE SOFTWARE. |
| 21 | * |
| 22 | */ |
| 23 | |
| 24 | #include <linux/perf_event.h> |
| 25 | #include <linux/init.h> |
| 26 | #include "amdgpu.h" |
| 27 | #include "amdgpu_pmu.h" |
| 28 | |
| 29 | #define PMU_NAME_SIZE 32 |
| 30 | #define NUM_FORMATS_AMDGPU_PMU 4 |
| 31 | #define NUM_FORMATS_DF_VEGA20 3 |
| 32 | #define NUM_EVENTS_DF_VEGA20 8 |
| 33 | #define NUM_EVENT_TYPES_VEGA20 1 |
| 34 | #define NUM_EVENTS_VEGA20_XGMI 2 |
| 35 | #define NUM_EVENTS_VEGA20_MAX NUM_EVENTS_VEGA20_XGMI |
| 36 | #define NUM_EVENT_TYPES_ARCTURUS 1 |
| 37 | #define NUM_EVENTS_ARCTURUS_XGMI 6 |
| 38 | #define NUM_EVENTS_ARCTURUS_MAX NUM_EVENTS_ARCTURUS_XGMI |
| 39 | |
| 40 | struct amdgpu_pmu_event_attribute { |
| 41 | struct device_attribute attr; |
| 42 | const char *event_str; |
| 43 | unsigned int type; |
| 44 | }; |
| 45 | |
| 46 | /* record to keep track of pmu entry per pmu type per device */ |
| 47 | struct amdgpu_pmu_entry { |
| 48 | struct list_head entry; |
| 49 | struct amdgpu_device *adev; |
| 50 | struct pmu pmu; |
| 51 | unsigned int pmu_perf_type; |
| 52 | char *pmu_type_name; |
| 53 | char *pmu_file_prefix; |
| 54 | struct attribute_group fmt_attr_group; |
| 55 | struct amdgpu_pmu_event_attribute *fmt_attr; |
| 56 | struct attribute_group evt_attr_group; |
| 57 | struct amdgpu_pmu_event_attribute *evt_attr; |
| 58 | }; |
| 59 | |
| 60 | static ssize_t amdgpu_pmu_event_show(struct device *dev, |
| 61 | struct device_attribute *attr, char *buf) |
| 62 | { |
| 63 | struct amdgpu_pmu_event_attribute *amdgpu_pmu_attr; |
| 64 | |
| 65 | amdgpu_pmu_attr = container_of(attr, struct amdgpu_pmu_event_attribute, |
| 66 | attr); |
| 67 | |
| 68 | if (!amdgpu_pmu_attr->type) |
| 69 | return sprintf(buf, fmt: "%s\n" , amdgpu_pmu_attr->event_str); |
| 70 | |
| 71 | return sprintf(buf, fmt: "%s,type=0x%x\n" , |
| 72 | amdgpu_pmu_attr->event_str, amdgpu_pmu_attr->type); |
| 73 | } |
| 74 | |
| 75 | static LIST_HEAD(amdgpu_pmu_list); |
| 76 | |
| 77 | |
| 78 | struct amdgpu_pmu_attr { |
| 79 | const char *name; |
| 80 | const char *config; |
| 81 | }; |
| 82 | |
| 83 | struct amdgpu_pmu_type { |
| 84 | const unsigned int type; |
| 85 | const unsigned int num_of_type; |
| 86 | }; |
| 87 | |
| 88 | struct amdgpu_pmu_config { |
| 89 | struct amdgpu_pmu_attr *formats; |
| 90 | unsigned int num_formats; |
| 91 | struct amdgpu_pmu_attr *events; |
| 92 | unsigned int num_events; |
| 93 | struct amdgpu_pmu_type *types; |
| 94 | unsigned int num_types; |
| 95 | }; |
| 96 | |
| 97 | /* |
| 98 | * Events fall under two categories: |
| 99 | * - PMU typed |
| 100 | * Events in /sys/bus/event_source/devices/amdgpu_<pmu_type>_<dev_num> have |
| 101 | * performance counter operations handled by one IP <pmu_type>. Formats and |
| 102 | * events should be defined by <pmu_type>_<asic_type>_formats and |
| 103 | * <pmu_type>_<asic_type>_events respectively. |
| 104 | * |
| 105 | * - Event config typed |
| 106 | * Events in /sys/bus/event_source/devices/amdgpu_<dev_num> have performance |
| 107 | * counter operations that can be handled by multiple IPs dictated by their |
| 108 | * "type" format field. Formats and events should be defined by |
| 109 | * amdgpu_pmu_formats and <asic_type>_events respectively. Format field |
| 110 | * "type" is generated in amdgpu_pmu_event_show and defined in |
| 111 | * <asic_type>_event_config_types. |
| 112 | */ |
| 113 | |
| 114 | static struct amdgpu_pmu_attr amdgpu_pmu_formats[NUM_FORMATS_AMDGPU_PMU] = { |
| 115 | { .name = "event" , .config = "config:0-7" }, |
| 116 | { .name = "instance" , .config = "config:8-15" }, |
| 117 | { .name = "umask" , .config = "config:16-23" }, |
| 118 | { .name = "type" , .config = "config:56-63" } |
| 119 | }; |
| 120 | |
| 121 | /* Vega20 events */ |
| 122 | static struct amdgpu_pmu_attr vega20_events[NUM_EVENTS_VEGA20_MAX] = { |
| 123 | { .name = "xgmi_link0_data_outbound" , |
| 124 | .config = "event=0x7,instance=0x46,umask=0x2" }, |
| 125 | { .name = "xgmi_link1_data_outbound" , |
| 126 | .config = "event=0x7,instance=0x47,umask=0x2" } |
| 127 | }; |
| 128 | |
| 129 | static struct amdgpu_pmu_type vega20_types[NUM_EVENT_TYPES_VEGA20] = { |
| 130 | { .type = AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI, |
| 131 | .num_of_type = NUM_EVENTS_VEGA20_XGMI } |
| 132 | }; |
| 133 | |
| 134 | static struct amdgpu_pmu_config vega20_config = { |
| 135 | .formats = amdgpu_pmu_formats, |
| 136 | .num_formats = ARRAY_SIZE(amdgpu_pmu_formats), |
| 137 | .events = vega20_events, |
| 138 | .num_events = ARRAY_SIZE(vega20_events), |
| 139 | .types = vega20_types, |
| 140 | .num_types = ARRAY_SIZE(vega20_types) |
| 141 | }; |
| 142 | |
| 143 | /* Vega20 data fabric (DF) events */ |
| 144 | static struct amdgpu_pmu_attr df_vega20_formats[NUM_FORMATS_DF_VEGA20] = { |
| 145 | { .name = "event" , .config = "config:0-7" }, |
| 146 | { .name = "instance" , .config = "config:8-15" }, |
| 147 | { .name = "umask" , .config = "config:16-23" } |
| 148 | }; |
| 149 | |
| 150 | static struct amdgpu_pmu_attr df_vega20_events[NUM_EVENTS_DF_VEGA20] = { |
| 151 | { .name = "cake0_pcsout_txdata" , |
| 152 | .config = "event=0x7,instance=0x46,umask=0x2" }, |
| 153 | { .name = "cake1_pcsout_txdata" , |
| 154 | .config = "event=0x7,instance=0x47,umask=0x2" }, |
| 155 | { .name = "cake0_pcsout_txmeta" , |
| 156 | .config = "event=0x7,instance=0x46,umask=0x4" }, |
| 157 | { .name = "cake1_pcsout_txmeta" , |
| 158 | .config = "event=0x7,instance=0x47,umask=0x4" }, |
| 159 | { .name = "cake0_ftiinstat_reqalloc" , |
| 160 | .config = "event=0xb,instance=0x46,umask=0x4" }, |
| 161 | { .name = "cake1_ftiinstat_reqalloc" , |
| 162 | .config = "event=0xb,instance=0x47,umask=0x4" }, |
| 163 | { .name = "cake0_ftiinstat_rspalloc" , |
| 164 | .config = "event=0xb,instance=0x46,umask=0x8" }, |
| 165 | { .name = "cake1_ftiinstat_rspalloc" , |
| 166 | .config = "event=0xb,instance=0x47,umask=0x8" } |
| 167 | }; |
| 168 | |
| 169 | static struct amdgpu_pmu_config df_vega20_config = { |
| 170 | .formats = df_vega20_formats, |
| 171 | .num_formats = ARRAY_SIZE(df_vega20_formats), |
| 172 | .events = df_vega20_events, |
| 173 | .num_events = ARRAY_SIZE(df_vega20_events), |
| 174 | .types = NULL, |
| 175 | .num_types = 0 |
| 176 | }; |
| 177 | |
| 178 | /* Arcturus events */ |
| 179 | static struct amdgpu_pmu_attr arcturus_events[NUM_EVENTS_ARCTURUS_MAX] = { |
| 180 | { .name = "xgmi_link0_data_outbound" , |
| 181 | .config = "event=0x7,instance=0x4b,umask=0x2" }, |
| 182 | { .name = "xgmi_link1_data_outbound" , |
| 183 | .config = "event=0x7,instance=0x4c,umask=0x2" }, |
| 184 | { .name = "xgmi_link2_data_outbound" , |
| 185 | .config = "event=0x7,instance=0x4d,umask=0x2" }, |
| 186 | { .name = "xgmi_link3_data_outbound" , |
| 187 | .config = "event=0x7,instance=0x4e,umask=0x2" }, |
| 188 | { .name = "xgmi_link4_data_outbound" , |
| 189 | .config = "event=0x7,instance=0x4f,umask=0x2" }, |
| 190 | { .name = "xgmi_link5_data_outbound" , |
| 191 | .config = "event=0x7,instance=0x50,umask=0x2" } |
| 192 | }; |
| 193 | |
| 194 | static struct amdgpu_pmu_type arcturus_types[NUM_EVENT_TYPES_ARCTURUS] = { |
| 195 | { .type = AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI, |
| 196 | .num_of_type = NUM_EVENTS_ARCTURUS_XGMI } |
| 197 | }; |
| 198 | |
| 199 | static struct amdgpu_pmu_config arcturus_config = { |
| 200 | .formats = amdgpu_pmu_formats, |
| 201 | .num_formats = ARRAY_SIZE(amdgpu_pmu_formats), |
| 202 | .events = arcturus_events, |
| 203 | .num_events = ARRAY_SIZE(arcturus_events), |
| 204 | .types = arcturus_types, |
| 205 | .num_types = ARRAY_SIZE(arcturus_types) |
| 206 | }; |
| 207 | |
| 208 | /* initialize perf counter */ |
| 209 | static int amdgpu_perf_event_init(struct perf_event *event) |
| 210 | { |
| 211 | struct hw_perf_event *hwc = &event->hw; |
| 212 | |
| 213 | /* test the event attr type check for PMU enumeration */ |
| 214 | if (event->attr.type != event->pmu->type) |
| 215 | return -ENOENT; |
| 216 | |
| 217 | /* update the hw_perf_event struct with config data */ |
| 218 | hwc->config = event->attr.config; |
| 219 | hwc->config_base = AMDGPU_PMU_PERF_TYPE_NONE; |
| 220 | |
| 221 | return 0; |
| 222 | } |
| 223 | |
| 224 | /* start perf counter */ |
| 225 | static void amdgpu_perf_start(struct perf_event *event, int flags) |
| 226 | { |
| 227 | struct hw_perf_event *hwc = &event->hw; |
| 228 | struct amdgpu_pmu_entry *pe = container_of(event->pmu, |
| 229 | struct amdgpu_pmu_entry, |
| 230 | pmu); |
| 231 | int target_cntr = 0; |
| 232 | |
| 233 | if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED))) |
| 234 | return; |
| 235 | |
| 236 | if ((!pe->adev->df.funcs) || |
| 237 | (!pe->adev->df.funcs->pmc_start)) |
| 238 | return; |
| 239 | |
| 240 | WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE)); |
| 241 | hwc->state = 0; |
| 242 | |
| 243 | switch (hwc->config_base) { |
| 244 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF: |
| 245 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI: |
| 246 | if (!(flags & PERF_EF_RELOAD)) { |
| 247 | target_cntr = pe->adev->df.funcs->pmc_start(pe->adev, |
| 248 | hwc->config, 0 /* unused */, |
| 249 | 1 /* add counter */); |
| 250 | if (target_cntr < 0) |
| 251 | break; |
| 252 | |
| 253 | hwc->idx = target_cntr; |
| 254 | } |
| 255 | |
| 256 | pe->adev->df.funcs->pmc_start(pe->adev, hwc->config, |
| 257 | hwc->idx, 0); |
| 258 | break; |
| 259 | default: |
| 260 | break; |
| 261 | } |
| 262 | |
| 263 | perf_event_update_userpage(event); |
| 264 | } |
| 265 | |
| 266 | /* read perf counter */ |
| 267 | static void amdgpu_perf_read(struct perf_event *event) |
| 268 | { |
| 269 | struct hw_perf_event *hwc = &event->hw; |
| 270 | struct amdgpu_pmu_entry *pe = container_of(event->pmu, |
| 271 | struct amdgpu_pmu_entry, |
| 272 | pmu); |
| 273 | u64 count, prev; |
| 274 | |
| 275 | if ((!pe->adev->df.funcs) || |
| 276 | (!pe->adev->df.funcs->pmc_get_count)) |
| 277 | return; |
| 278 | |
| 279 | prev = local64_read(&hwc->prev_count); |
| 280 | do { |
| 281 | switch (hwc->config_base) { |
| 282 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF: |
| 283 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI: |
| 284 | pe->adev->df.funcs->pmc_get_count(pe->adev, |
| 285 | hwc->config, hwc->idx, &count); |
| 286 | break; |
| 287 | default: |
| 288 | count = 0; |
| 289 | break; |
| 290 | } |
| 291 | } while (!local64_try_cmpxchg(l: &hwc->prev_count, old: &prev, new: count)); |
| 292 | |
| 293 | local64_add(count - prev, &event->count); |
| 294 | } |
| 295 | |
| 296 | /* stop perf counter */ |
| 297 | static void amdgpu_perf_stop(struct perf_event *event, int flags) |
| 298 | { |
| 299 | struct hw_perf_event *hwc = &event->hw; |
| 300 | struct amdgpu_pmu_entry *pe = container_of(event->pmu, |
| 301 | struct amdgpu_pmu_entry, |
| 302 | pmu); |
| 303 | |
| 304 | if (hwc->state & PERF_HES_UPTODATE) |
| 305 | return; |
| 306 | |
| 307 | if ((!pe->adev->df.funcs) || |
| 308 | (!pe->adev->df.funcs->pmc_stop)) |
| 309 | return; |
| 310 | |
| 311 | switch (hwc->config_base) { |
| 312 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF: |
| 313 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI: |
| 314 | pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc->idx, |
| 315 | 0); |
| 316 | break; |
| 317 | default: |
| 318 | break; |
| 319 | } |
| 320 | |
| 321 | WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED); |
| 322 | hwc->state |= PERF_HES_STOPPED; |
| 323 | |
| 324 | if (hwc->state & PERF_HES_UPTODATE) |
| 325 | return; |
| 326 | |
| 327 | amdgpu_perf_read(event); |
| 328 | hwc->state |= PERF_HES_UPTODATE; |
| 329 | } |
| 330 | |
| 331 | /* add perf counter */ |
| 332 | static int amdgpu_perf_add(struct perf_event *event, int flags) |
| 333 | { |
| 334 | struct hw_perf_event *hwc = &event->hw; |
| 335 | int retval = 0, target_cntr; |
| 336 | struct amdgpu_pmu_entry *pe = container_of(event->pmu, |
| 337 | struct amdgpu_pmu_entry, |
| 338 | pmu); |
| 339 | |
| 340 | if ((!pe->adev->df.funcs) || |
| 341 | (!pe->adev->df.funcs->pmc_start)) |
| 342 | return -EINVAL; |
| 343 | |
| 344 | switch (pe->pmu_perf_type) { |
| 345 | case AMDGPU_PMU_PERF_TYPE_DF: |
| 346 | hwc->config_base = AMDGPU_PMU_EVENT_CONFIG_TYPE_DF; |
| 347 | break; |
| 348 | case AMDGPU_PMU_PERF_TYPE_ALL: |
| 349 | hwc->config_base = (hwc->config >> |
| 350 | AMDGPU_PMU_EVENT_CONFIG_TYPE_SHIFT) & |
| 351 | AMDGPU_PMU_EVENT_CONFIG_TYPE_MASK; |
| 352 | break; |
| 353 | } |
| 354 | |
| 355 | event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED; |
| 356 | |
| 357 | switch (hwc->config_base) { |
| 358 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF: |
| 359 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI: |
| 360 | target_cntr = pe->adev->df.funcs->pmc_start(pe->adev, |
| 361 | hwc->config, 0 /* unused */, |
| 362 | 1 /* add counter */); |
| 363 | if (target_cntr < 0) |
| 364 | retval = target_cntr; |
| 365 | else |
| 366 | hwc->idx = target_cntr; |
| 367 | |
| 368 | break; |
| 369 | default: |
| 370 | return 0; |
| 371 | } |
| 372 | |
| 373 | if (retval) |
| 374 | return retval; |
| 375 | |
| 376 | if (flags & PERF_EF_START) |
| 377 | amdgpu_perf_start(event, PERF_EF_RELOAD); |
| 378 | |
| 379 | return retval; |
| 380 | } |
| 381 | |
| 382 | /* delete perf counter */ |
| 383 | static void amdgpu_perf_del(struct perf_event *event, int flags) |
| 384 | { |
| 385 | struct hw_perf_event *hwc = &event->hw; |
| 386 | struct amdgpu_pmu_entry *pe = container_of(event->pmu, |
| 387 | struct amdgpu_pmu_entry, |
| 388 | pmu); |
| 389 | if ((!pe->adev->df.funcs) || |
| 390 | (!pe->adev->df.funcs->pmc_stop)) |
| 391 | return; |
| 392 | |
| 393 | amdgpu_perf_stop(event, PERF_EF_UPDATE); |
| 394 | |
| 395 | switch (hwc->config_base) { |
| 396 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_DF: |
| 397 | case AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI: |
| 398 | pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc->idx, |
| 399 | 1); |
| 400 | break; |
| 401 | default: |
| 402 | break; |
| 403 | } |
| 404 | |
| 405 | perf_event_update_userpage(event); |
| 406 | } |
| 407 | |
| 408 | static void amdgpu_pmu_create_event_attrs_by_type( |
| 409 | struct attribute_group *attr_group, |
| 410 | struct amdgpu_pmu_event_attribute *pmu_attr, |
| 411 | struct amdgpu_pmu_attr events[], |
| 412 | int s_offset, |
| 413 | int e_offset, |
| 414 | unsigned int type) |
| 415 | { |
| 416 | int i; |
| 417 | |
| 418 | pmu_attr += s_offset; |
| 419 | |
| 420 | for (i = s_offset; i < e_offset; i++) { |
| 421 | attr_group->attrs[i] = &pmu_attr->attr.attr; |
| 422 | sysfs_attr_init(&pmu_attr->attr.attr); |
| 423 | pmu_attr->attr.attr.name = events[i].name; |
| 424 | pmu_attr->attr.attr.mode = 0444; |
| 425 | pmu_attr->attr.show = amdgpu_pmu_event_show; |
| 426 | pmu_attr->event_str = events[i].config; |
| 427 | pmu_attr->type = type; |
| 428 | pmu_attr++; |
| 429 | } |
| 430 | } |
| 431 | |
| 432 | static void amdgpu_pmu_create_attrs(struct attribute_group *attr_group, |
| 433 | struct amdgpu_pmu_event_attribute *pmu_attr, |
| 434 | struct amdgpu_pmu_attr events[], |
| 435 | int num_events) |
| 436 | { |
| 437 | amdgpu_pmu_create_event_attrs_by_type(attr_group, pmu_attr, events, s_offset: 0, |
| 438 | e_offset: num_events, type: AMDGPU_PMU_EVENT_CONFIG_TYPE_NONE); |
| 439 | } |
| 440 | |
| 441 | |
| 442 | static int amdgpu_pmu_alloc_pmu_attrs( |
| 443 | struct attribute_group *fmt_attr_group, |
| 444 | struct amdgpu_pmu_event_attribute **fmt_attr, |
| 445 | struct attribute_group *evt_attr_group, |
| 446 | struct amdgpu_pmu_event_attribute **evt_attr, |
| 447 | struct amdgpu_pmu_config *config) |
| 448 | { |
| 449 | *fmt_attr = kcalloc(config->num_formats, sizeof(**fmt_attr), |
| 450 | GFP_KERNEL); |
| 451 | |
| 452 | if (!(*fmt_attr)) |
| 453 | return -ENOMEM; |
| 454 | |
| 455 | fmt_attr_group->attrs = kcalloc(config->num_formats + 1, |
| 456 | sizeof(*fmt_attr_group->attrs), GFP_KERNEL); |
| 457 | |
| 458 | if (!fmt_attr_group->attrs) |
| 459 | goto err_fmt_attr_grp; |
| 460 | |
| 461 | *evt_attr = kcalloc(config->num_events, sizeof(**evt_attr), GFP_KERNEL); |
| 462 | |
| 463 | if (!(*evt_attr)) |
| 464 | goto err_evt_attr; |
| 465 | |
| 466 | evt_attr_group->attrs = kcalloc(config->num_events + 1, |
| 467 | sizeof(*evt_attr_group->attrs), GFP_KERNEL); |
| 468 | |
| 469 | if (!evt_attr_group->attrs) |
| 470 | goto err_evt_attr_grp; |
| 471 | |
| 472 | return 0; |
| 473 | err_evt_attr_grp: |
| 474 | kfree(objp: *evt_attr); |
| 475 | err_evt_attr: |
| 476 | kfree(objp: fmt_attr_group->attrs); |
| 477 | err_fmt_attr_grp: |
| 478 | kfree(objp: *fmt_attr); |
| 479 | return -ENOMEM; |
| 480 | } |
| 481 | |
| 482 | /* init pmu tracking per pmu type */ |
| 483 | static int init_pmu_entry_by_type_and_add(struct amdgpu_pmu_entry *pmu_entry, |
| 484 | struct amdgpu_pmu_config *config) |
| 485 | { |
| 486 | const struct attribute_group *attr_groups[] = { |
| 487 | &pmu_entry->fmt_attr_group, |
| 488 | &pmu_entry->evt_attr_group, |
| 489 | NULL |
| 490 | }; |
| 491 | char pmu_name[PMU_NAME_SIZE]; |
| 492 | int ret = 0, total_num_events = 0; |
| 493 | |
| 494 | pmu_entry->pmu = (struct pmu){ |
| 495 | .event_init = amdgpu_perf_event_init, |
| 496 | .add = amdgpu_perf_add, |
| 497 | .del = amdgpu_perf_del, |
| 498 | .start = amdgpu_perf_start, |
| 499 | .stop = amdgpu_perf_stop, |
| 500 | .read = amdgpu_perf_read, |
| 501 | .task_ctx_nr = perf_invalid_context, |
| 502 | }; |
| 503 | |
| 504 | ret = amdgpu_pmu_alloc_pmu_attrs(fmt_attr_group: &pmu_entry->fmt_attr_group, |
| 505 | fmt_attr: &pmu_entry->fmt_attr, |
| 506 | evt_attr_group: &pmu_entry->evt_attr_group, |
| 507 | evt_attr: &pmu_entry->evt_attr, |
| 508 | config); |
| 509 | |
| 510 | if (ret) |
| 511 | goto err_out; |
| 512 | |
| 513 | amdgpu_pmu_create_attrs(attr_group: &pmu_entry->fmt_attr_group, pmu_attr: pmu_entry->fmt_attr, |
| 514 | events: config->formats, num_events: config->num_formats); |
| 515 | |
| 516 | if (pmu_entry->pmu_perf_type == AMDGPU_PMU_PERF_TYPE_ALL) { |
| 517 | int i; |
| 518 | |
| 519 | for (i = 0; i < config->num_types; i++) { |
| 520 | amdgpu_pmu_create_event_attrs_by_type( |
| 521 | attr_group: &pmu_entry->evt_attr_group, |
| 522 | pmu_attr: pmu_entry->evt_attr, |
| 523 | events: config->events, |
| 524 | s_offset: total_num_events, |
| 525 | e_offset: total_num_events + |
| 526 | config->types[i].num_of_type, |
| 527 | type: config->types[i].type); |
| 528 | total_num_events += config->types[i].num_of_type; |
| 529 | } |
| 530 | } else { |
| 531 | amdgpu_pmu_create_attrs(attr_group: &pmu_entry->evt_attr_group, |
| 532 | pmu_attr: pmu_entry->evt_attr, |
| 533 | events: config->events, num_events: config->num_events); |
| 534 | total_num_events = config->num_events; |
| 535 | } |
| 536 | |
| 537 | pmu_entry->pmu.attr_groups = kmemdup(attr_groups, sizeof(attr_groups), |
| 538 | GFP_KERNEL); |
| 539 | |
| 540 | if (!pmu_entry->pmu.attr_groups) { |
| 541 | ret = -ENOMEM; |
| 542 | goto err_attr_group; |
| 543 | } |
| 544 | |
| 545 | snprintf(buf: pmu_name, PMU_NAME_SIZE, fmt: "%s_%d" , pmu_entry->pmu_file_prefix, |
| 546 | adev_to_drm(adev: pmu_entry->adev)->primary->index); |
| 547 | |
| 548 | ret = perf_pmu_register(pmu: &pmu_entry->pmu, name: pmu_name, type: -1); |
| 549 | |
| 550 | if (ret) |
| 551 | goto err_register; |
| 552 | |
| 553 | if (pmu_entry->pmu_perf_type != AMDGPU_PMU_PERF_TYPE_ALL) |
| 554 | pr_info("Detected AMDGPU %s Counters. # of Counters = %d.\n" , |
| 555 | pmu_entry->pmu_type_name, total_num_events); |
| 556 | else |
| 557 | pr_info("Detected AMDGPU %d Perf Events.\n" , total_num_events); |
| 558 | |
| 559 | |
| 560 | list_add_tail(new: &pmu_entry->entry, head: &amdgpu_pmu_list); |
| 561 | |
| 562 | return 0; |
| 563 | err_register: |
| 564 | kfree(objp: pmu_entry->pmu.attr_groups); |
| 565 | err_attr_group: |
| 566 | kfree(objp: pmu_entry->fmt_attr_group.attrs); |
| 567 | kfree(objp: pmu_entry->fmt_attr); |
| 568 | kfree(objp: pmu_entry->evt_attr_group.attrs); |
| 569 | kfree(objp: pmu_entry->evt_attr); |
| 570 | err_out: |
| 571 | pr_warn("Error initializing AMDGPU %s PMUs.\n" , |
| 572 | pmu_entry->pmu_type_name); |
| 573 | return ret; |
| 574 | } |
| 575 | |
| 576 | /* destroy all pmu data associated with target device */ |
| 577 | void amdgpu_pmu_fini(struct amdgpu_device *adev) |
| 578 | { |
| 579 | struct amdgpu_pmu_entry *pe, *temp; |
| 580 | |
| 581 | list_for_each_entry_safe(pe, temp, &amdgpu_pmu_list, entry) { |
| 582 | if (pe->adev != adev) |
| 583 | continue; |
| 584 | list_del(entry: &pe->entry); |
| 585 | perf_pmu_unregister(pmu: &pe->pmu); |
| 586 | kfree(objp: pe->pmu.attr_groups); |
| 587 | kfree(objp: pe->fmt_attr_group.attrs); |
| 588 | kfree(objp: pe->fmt_attr); |
| 589 | kfree(objp: pe->evt_attr_group.attrs); |
| 590 | kfree(objp: pe->evt_attr); |
| 591 | kfree(objp: pe); |
| 592 | } |
| 593 | } |
| 594 | |
| 595 | static struct amdgpu_pmu_entry *create_pmu_entry(struct amdgpu_device *adev, |
| 596 | unsigned int pmu_type, |
| 597 | char *pmu_type_name, |
| 598 | char *pmu_file_prefix) |
| 599 | { |
| 600 | struct amdgpu_pmu_entry *pmu_entry; |
| 601 | |
| 602 | pmu_entry = kzalloc(sizeof(struct amdgpu_pmu_entry), GFP_KERNEL); |
| 603 | |
| 604 | if (!pmu_entry) |
| 605 | return pmu_entry; |
| 606 | |
| 607 | pmu_entry->adev = adev; |
| 608 | pmu_entry->fmt_attr_group.name = "format" ; |
| 609 | pmu_entry->fmt_attr_group.attrs = NULL; |
| 610 | pmu_entry->evt_attr_group.name = "events" ; |
| 611 | pmu_entry->evt_attr_group.attrs = NULL; |
| 612 | pmu_entry->pmu_perf_type = pmu_type; |
| 613 | pmu_entry->pmu_type_name = pmu_type_name; |
| 614 | pmu_entry->pmu_file_prefix = pmu_file_prefix; |
| 615 | |
| 616 | return pmu_entry; |
| 617 | } |
| 618 | |
| 619 | /* init amdgpu_pmu */ |
| 620 | int amdgpu_pmu_init(struct amdgpu_device *adev) |
| 621 | { |
| 622 | int ret = 0; |
| 623 | struct amdgpu_pmu_entry *pmu_entry, *pmu_entry_df; |
| 624 | |
| 625 | switch (adev->asic_type) { |
| 626 | case CHIP_VEGA20: |
| 627 | pmu_entry_df = create_pmu_entry(adev, pmu_type: AMDGPU_PMU_PERF_TYPE_DF, |
| 628 | pmu_type_name: "DF" , pmu_file_prefix: "amdgpu_df" ); |
| 629 | |
| 630 | if (!pmu_entry_df) |
| 631 | return -ENOMEM; |
| 632 | |
| 633 | ret = init_pmu_entry_by_type_and_add(pmu_entry: pmu_entry_df, |
| 634 | config: &df_vega20_config); |
| 635 | |
| 636 | if (ret) { |
| 637 | kfree(objp: pmu_entry_df); |
| 638 | return ret; |
| 639 | } |
| 640 | |
| 641 | pmu_entry = create_pmu_entry(adev, pmu_type: AMDGPU_PMU_PERF_TYPE_ALL, |
| 642 | pmu_type_name: "" , pmu_file_prefix: "amdgpu" ); |
| 643 | |
| 644 | if (!pmu_entry) { |
| 645 | amdgpu_pmu_fini(adev); |
| 646 | return -ENOMEM; |
| 647 | } |
| 648 | |
| 649 | ret = init_pmu_entry_by_type_and_add(pmu_entry, |
| 650 | config: &vega20_config); |
| 651 | |
| 652 | if (ret) { |
| 653 | kfree(objp: pmu_entry); |
| 654 | amdgpu_pmu_fini(adev); |
| 655 | return ret; |
| 656 | } |
| 657 | |
| 658 | break; |
| 659 | case CHIP_ARCTURUS: |
| 660 | pmu_entry = create_pmu_entry(adev, pmu_type: AMDGPU_PMU_PERF_TYPE_ALL, |
| 661 | pmu_type_name: "" , pmu_file_prefix: "amdgpu" ); |
| 662 | if (!pmu_entry) |
| 663 | return -ENOMEM; |
| 664 | |
| 665 | ret = init_pmu_entry_by_type_and_add(pmu_entry, |
| 666 | config: &arcturus_config); |
| 667 | |
| 668 | if (ret) { |
| 669 | kfree(objp: pmu_entry); |
| 670 | return -ENOMEM; |
| 671 | } |
| 672 | |
| 673 | break; |
| 674 | |
| 675 | default: |
| 676 | return 0; |
| 677 | } |
| 678 | |
| 679 | return ret; |
| 680 | } |
| 681 | |